Reproduce “SynAct: A Reasoning-Acting Large Language Model Agent for Adaptive Synthesis Optimization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Fangzhou Liu, Peiyi Han, Jiawei Liu, Yuan Pu, Zhuolun He, Rongliang Fu, Tsung-Yi Ho, Bei Yu
SynAct introduces an adaptive, closed-loop LLM agent for optimizing logic synthesis, which iteratively diagnoses PPA reports and issues targeted commands within commercial tools. The method achieved superior timing improvement, reducing average Worst Negative Slack to 27.03% and Total Negative Slack to 17.86% of bootstrap results on OpenCores designs, while balancing area and power metrics.
Key insights
- Closed-loop, iterative interaction allows the LLM agent to dynamically adapt optimization strategies based on real-time PPA feedback, overcoming the limitations of static script generation.
- A multi-layer GraphRAG module effectively prunes extensive documentation, providing precise, scenario-relevant knowledge to the LLM, thereby mitigating information overload and improving command generation accuracy.
- Systematic reuse of historical optimization experience through Bayesian optimization within a GrammarVAE latent space efficiently guides exploration, preventing redundant efforts and focusing on effective strategies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.